Breast Lump Detection and Localization with a Tactile Glove Using Deep Learning

TL;DR

This paper tackles the need for radiation-free, accessible breast lump screening by combining a fabric-based tactile glove with deep learning to detect and localize lumps in silicone breast prototypes. The authors fabricate SBPs, collect a dataset from ten naive participants and an oncologist, and benchmark multiple time-series DL architectures, with InceptionTime providing the best lump-detection performance (~$94\%$). Through within-user data splits and incremental transfer learning using oncologist data, lump-presence accuracy reaches ~$95.0\%$, while lump-size and location accuracies reach ~${88.5}\%$ and ${82.9}\%$, respectively. The study demonstrates that the tactile glove–DL approach can augment palpation, potentially enabling more frequent checks by non-experts and healthcare providers, though real-breast applicability and generalization to diverse patient anatomies remain to be validated.

Abstract

Breast cancer is the leading cause of mortality among women. Inspection of breasts by palpation is the key to early detection. We aim to create a wearable tactile glove that could localize the lump in breasts using deep learning (DL). In this work, we present our flexible fabric-based and soft wearable tactile glove for detecting the lumps within custom-made silicone breast prototypes (SBPs). SBPs are made of soft silicone that imitates the human skin and the inner part of the breast. Ball-shaped silicone tumors of 1.5-, 1.75- and 2.0-cm diameters are embedded inside to create another set with lumps. Our approach is based on the InceptionTime DL architecture with transfer learning between experienced and non-experienced users. We collected a dataset from 10 naive participants and one oncologist-mammologist palpating SBPs. We demonstrated that the DL model can classify lump presence, size and location with an accuracy of 82.22%, 67.08% and 62.63%, respectively. In addition, we showed that the model adapted to unseen experienced users with an accuracy of 95.01%, 88.54% and 82.98% for lump presence, size and location classification, respectively. This technology can assist inexperienced users or healthcare providers, thus facilitating more frequent routine checks.

Figures (9)

• Figure 1: a) The experimental scenario of the Tactile Glove for palpation of the breast with a lump, and b) side view of the breast structure with a lump.
• Figure 2: Tactile glove with 30 pressure sensors and two accelerometers (a) top view and (b) side view.
• Figure 3: Fabrication pipeline of the silicone breast prototype. It shows the step-by-step SBP casting: (1) and (2) - skin formation by pressing upper mold on lower mold, (3) and (5) - inner substance casting, (4) - lump formation and placement in between inner substance (skip this step to create a prototype without a lump), and (6) - the creation of the thin disk.
• Figure 4: Schematics of the learning approach. UL -- user-level, WU -- within-user, STL -- single-task learning, MTL -- multi-task learning.
• Figure 5: Experimental setup for the lump examination in the silicone breast prototypes. Tactile glove incorporates 3-axes accelerometers (ADXL) and barometers (TakkStrip2), a microcontroller board (STM32) and a multiplexer (TakkFast).